iva-agent

• Voice — voice, audio and video notes transcribed with Deepgram nova-3; auto-detects ru/uz/en.
• Vision — photos described by your provider's own vision model; no extra key, no extra bill.
• Rich replies — tables, checklists, collapsible blocks and formulas render natively in Telegram via Bot API 10.1 rich messages; plain formatting keeps its proven path, with a graceful fallback.
• Quiet update checks — once a day Iva checks for a newer stable release without spending model tokens. If one exists, Telegram offers Update or Later once; otherwise it says nothing.
• Layered memory — remembers across months, long after the chat window has scrolled away.
• Personal CRM — who your people are, what you agreed, when to follow up.
• Search by meaning — BM25 plus link-graph rerank, any language; optional vector mode with one key.
• Decision cards — what you chose, when and why; old versions stay in a dated History.
• Tasks & reminders — priorities, due dates and a morning digest.
• Web search — four pluggable providers: Tavily, Exa, Parallel or Brave.
• Google Workspace — Gmail, Calendar, Drive, Sheets, Docs and Tasks from chat via the gws CLI; installed for you, with a guided key setup right in the conversation.
• Skills & MCP — drop one file to add a procedure or connect an MCP server; keys stay in .env.
• Personal Telegram — userbot (beta) — read and send from your own account, not just the bot; connect by chat (QR, no terminal). Rough and buggy — opt-in, at your own risk. A server-side anti-ban guardrail (FloodWait compliance + randomized pacing + circuit-breaker) is enforced, not just advised. Details.
• Safe to forward — forwarded text, captions and voice transcripts pass an injection screen before the model reads them. A flagged message or transcript reaches the model tagged as data rather than as an instruction; for media captions the screen runs but the tag does not travel with it yet.
• Token accounting — every model step is logged; /usage reports it for free.

git clone https://github.com/smixs/iva-agent.git ~/iva
cd ~/iva && bash install.sh

The installer reuses the existing checkout instead of re-cloning, keeps .env and the vault untouched, and installs the same dependencies. A fork or a branch works through variables read at startup: REPO_URL=…, BRANCH=…, INSTALL_DIR=… (defaults: this repo, main, ~/iva). Details: docs/install.md.

19.08.2026

v0.3.28

• The update cleans up after itself: the disk keeps the running version and one rollback, everything else goes. Before a build only the running version stays — the build takes the rollback slot, so the peak is two versions (~400 MB each), not three. In the finish, removing old versions is the first step, ahead of npm i -g @googleworkspace/cli@latest, which downloads onto the same disk; a failed cleanup chore no longer cancels it. iva doctor removes leftovers of interrupted builds and surplus versions by itself, under the update lock, and prints versions on disk: N (current …, rollback …) — X GB free; with an update running it says version cleanup skipped, with a corrupt active.json it reports and deletes nothing. The cleanup before a build starts working from the update after this one — iva update runs on the installed CLI. Prompted by a test droplet with 8.7 GB that stalled on three copies of node_modules (#194). New troubleshooting section «Disk full during update», an honest disk line in install.
• The promoted runtime carries every source tree: the stable runtime snapshot copied agent/ and scripts/ by hand, packages/ never reached it, and on an install with a non-empty custom layer the agent died with [UNRESOLVED_IMPORT] on agent/lib/data-dir.ts in a restart loop. One list now — RUNTIME_SOURCE_TREES — feeds both the snapshot and the replica smoke, the layout digest moves to v3 so a runtime staged without packages/ rebuilds itself, and a guard test resolves every import of those trees and fails on a tree the snapshot would leave behind (#192, #191).
• Timezones come back in canonical spelling: Intl accepted europe/moscow in .env, systemd rejected OnCalendar=*-*-* 05:00:00 europe/moscow, and the nightly timer never came up. The validator now returns what Intl resolves — Europe/Moscow; aliases canonicalize (US/Pacific → America/Los_Angeles); property tests pin idempotence and insensitivity to case and padding (#190).
• A vision model per provider: instead of a constant in code, a variable in .env each, blank meaning the provider default — OLLAMA_VISION_MODEL (gemma4:31b), OPENCODE_VISION_MODEL (qwen3.7-plus), OPENROUTER_VISION_MODEL (google/gemini-2.5-flash); Codex has none, the subscription is multimodal. iva config asks for the vision model right after the text model and writes it next to *_MODEL. The OpenCode Go default comes from live runs on 18.08: gpt-5.6-luna answers 400 to any image, minimax-m3 wraps its answer in <think> inside content, qwen3.7-plus returns a clean description with OCR in 4–6 seconds.
• Three Telegram fixes: the Stop button only in private chats — its callback has been rejected in groups since the last release, and the button hung there dead. Messages buffered during a rolling update pass the inbound gate with the same warning to the model and the same [security] inbound flagged line as fresh ones — the verdict used to be ignored. Prose that looks like a code placeholder (the price is 50 dollars, tables with padded numbers) is no longer cut out of an HTML reply: the placeholder moved to the Private Use Area, such code points from outside are stripped first, and a seeded property test walks digits, space runs and code spans mixed in prose.

18.08.2026

v0.3.27

• MCP servers of a plugin, both transports: streamable-http and sse in mcp.json become a generated eve connection mcp-<name>--<server>, and ${VAR} in a header is filled at run time from data/custom/plugins/<name>.env, so no token is baked into a build. stdio runs as the systemd unit iva-mcp-<name>-<server>.service behind Iva's own MCP proxy (services/mcp-proxy/, @modelcontextprotocol/sdk): the agent reaches it over 127.0.0.1:<port>/mcp with a bearer, and the token lives in data/plugin-data/<name>/mcp-<server>.token at mode 0600. The server sees only PATH, HOME, PLUGIN_ROOT, PLUGIN_DATA, its own env from mcp.json and <name>.env — nothing of the agent's environment reaches it. A second switch joins the first: trusted, through iva plugin trust | untrust, and add prints the processes and asks Start these processes on this machine? [y/N] (--trust answers yes; a shell without a terminal answers no). Ports are handed out from 8730, once, and stay until the plugin is removed. Proven end to end against a real stdio server and a real client from the SDK, not fakes.
• Plugin services: sh.iva/services/<svc>/service.json with {command,args,port} becomes the unit iva-plugin-<name>-<svc>.service — env IVA_SERVICE_PORT, IVA_DATA_DIR, PLUGIN_ROOT and PLUGIN_DATA, the service's own folder as the working directory, started only while the plugin is enabled and trusted. iva plugin update restarts the units of a plugin whose content changed, iva update brings them back right after the flip, and iva doctor lists the units, prints is-active and calls GET /health on every MCP proxy. sh.iva/ is now two kinds: an eve Extension (sh.iva/package.json, built into a version) and services, which never rebuild one.
• The default Marketplace is live — smixs/iva-plugins: the list sits at github.com/smixs/iva-plugins and carries two plugins. trace is the Trace viewer: the schema of Iva with the path of a turn lit across it, the feed of turns, replay at ×1, ×2 and ×4, tiles for today, 7 and 14 days; it listens on loopback only, and iva trace open prints the ready ssh tunnel command. hello is the demo code plugin authors copy: one skill, one tool. Three commands to get there: iva plugin add trace, iva plugin trust trace, iva trace open. Checked live from the public list: list --available, add trace pinned to a sha, update, remove.
• Plugin docs: docs/plugins.md (Russian: docs/ru/plugins.md) — what a plugin is, how to install one (a folder, owner/repo[/subdir][@ref], a git URL or a name from a Marketplace), how enabled differs from trusted, what iva update does to plugins, how to write your own (skills, an Extension under sh.iva/, mcp.json, services) and what you risk; SECURITY.md gains a Plugins section. Also: the iva plugin CLI is split into modules with no command changed, and on a development checkout add and remove of a code plugin stop promising a build that never happened and say plainly that no version was built there.

17.08.2026

v0.3.26

• Plugins by name from a Marketplace: iva plugin add trace finds the plugin in an .agents/plugins/marketplace.json list (the Codex convention — a path string or local | url | git-subdir; npm and policy entries are skipped aloud) and installs it through the same path as a folder or a git URL. iva plugin marketplace add | remove | list manage your lists, iva plugin list --available shows what is on offer; the default list is smixs/iva-plugins. Lists are cached under data/marketplace-cache/; offline, the cache is used and marked as possibly stale. Only https://, ssh:// and git@host: are accepted — file:// and local paths from a foreign list are refused, and plugin git never waits for a password in the terminal. A name offered by two lists is refused with an add <name>@<list> hint.
• Plugin code is built into the version: a plugin with an eve Extension under sh.iva/ now rides the update rails — a copy inside the version, npm ci, eve extension build with the very eve that runs Iva, a mount under its own namespace, then probe, flip and restart. The plugin's config is read from data/custom/plugins/<name>.config.json at start, never baked into the build. A failed build on add | update | enable rolls the install back and leaves the running version alone; on iva update the version is built without that plugin, the plugin is switched off and one Alert says what to do. iva doctor reports whether each plugin's code is built into the current version. Verified against the real eve 0.30.8, not only fakes.
• iva trace reads the turn journal: iva trace tail streams events live (--since N prints the last N lines first), iva trace show lists the last 20 turns, iva trace show <turn_12 | tg:<chat>:<message> | session id | last> prints one turn with its steps, tools and subagent (--full without the cap, --json raw lines), iva trace open prints the viewer address and a ready ssh tunnel command. Control characters in content never reach the terminal. The contract in docs/trace.md now says how night turns stitch: inside one session, turns are split by turn_N, and a line without a turn key belongs to the most recent open turn of its session.
• iva post replaces the Python sender: a rich post to another allowlisted chat goes from the CLI through the Outbox and the outbound gate (ADR-0005) — the recipient is the digest chat or an explicit --chat from the allowlist, the token comes from .env, a local image reaches the public upload host only behind --allow-upload and only as a real media file by magic bytes, 50 media at most. A post that cannot render rich is not downgraded to HTML without its images: it fails with one Bot API call and the error; a turn reply keeps its fallback. The Python copies of the security checks are gone from the skill — the gate lives in agent/lib/security-gate.ts, only data (the pattern list and the secret-key inventory) stays beside the skill. Missed in the 0.3.26 notes.

v0.3.25

• Iva reads rich Telegram posts: longreads from the new editor (rich_message — headings, lists, tables, quotes, collages, media captions) used to arrive empty because their text field is empty and the content lives in blocks. Now the text is read whole and in order, photos and videos from any block go through the usual media pipeline (10 per message, the rest announced in one line), truncation is visible. Works everywhere text is read: single message, album, forward, quoted message, media caption. Ordinary messages are untouched. The gap was shown by contributor PR #189.
• Plugins install from the terminal: iva plugin add <folder | owner/repo[/subdir][@ref] | git URL> puts an Agent Plugins 1.0.0 folder under data/custom/plugins/, records it in data/custom/plugins.json, and its skills work from the next turn — no build, no restart. iva plugin list | remove | update | enable | disable | sync manage them; iva doctor gains a Plugins section. Only the owner, only the terminal: no Telegram command and no model tool by design (ADR-0009). Plugin code and MCP servers are read and reported but not built yet — that lands in the next patch.
• A turn journal: data/trace/YYYY-MM-DD.jsonl records every turn from the Bridge to the Outbox — allowlist, Gate verdicts, context parts, model steps, tool calls, subagents, delivery — one JSON line per event, kept 14 days, content capped and switchable off with captureContent in data/settings.json. The format is documented in docs/trace.md (ADR-0010); a terminal reader (iva trace) and a viewer plugin follow in the next patch.